An Introduction to Extreme Value Theory

Petra Friederichs
Meteorological Institute
University of Bonn

COPS Summer School, July/August, 2007

Applications of EVT
Finance
distribution of income has so called fat tails
value-at-risk: maximal daily lost
re-assurance
Hydrology
protection against flood
Q100: maximal flow that is expected once every 100 years
Meteorology
extreme winds
risk assessment (e.g. ICE, power plants)
heavy precipitation events
heat waves, hurricanes, droughts
extremes in a changing climate

Application
In classical statistics:
focus on AVERAGE behavior
of stochastic process
central limit theorem

In extreme value theory:

focus on extreme and rare
events
Fisher-Tippett theorem
What is an extreme?
from Ulrike Schneider

Extreme Value Theory

Block Maximum

M n=max {X 1 , , X n }
for n
M n follows a Generalized Extreme Value (GEV) distribution
Peak over Threshold (POT)

{ X i uX i u}
very large threshold u
follow a Generalized Pareto Distribution (GPD)
Poisson-Point GPD Process
combines POT with Poisson point process

Extreme Value Theory

Block Maximum

M n =max {X 1 , , X n }

Example: station precipitation (DWD) at Dresden

1948 2004: 57 seasonal (Nov-March and May-Sept)
(maxima over approximately 150 days)
Summer
Winter

Dresden

GEV Fisher-Tippett Theorem

The distribution of

M n =max {X 1 , , X n }

converges to ( n )
1 /

y
G y=exp [1
]

=0

y
G y=exp exp

which is called the Generalized Extreme Value (GEV) distribution.

It has three parameters

location parameter
scale parameter
shape parameter

GEV Types of Distributions

GEV has 3 types depending on shape parameter

=0
G x =expexp[x ]

Gumbel

Frchet

=1/0

x
G x =exp[1 ]

=1/0
x
G x =exp[1 ]

Weibull

y
x=

GEV Types of Distributions

GEV has 3 types depending on shape parameter
Gumbel

=0

exponential tail
Frchet

=1/0

so called fat tail

Weibull

=1/0

upper finite endpoint

y
x=

GEV Types of Distribution

Conditions to the sample {X 1 , , X n } from which the maxima are drawn

Xi

must be independently identically distributed (i.i.d.)

Let

F x be the distribution of X i

F x

is in the domain of attraction of a Gumbel type GEV iff

1F xtb x t
lim
=e
1F x
x
for all

p=0.99
p=0.9

F(x)

t0

Exponential decay in the tail of

F x
x0.9

x0.99

GEV Types of Distribution

Conditions to the sample {X 1 , , X n } from which the maxima are drawn

Xi

must be independently identically distributed (i.i.d.)

Let

F x be the distribution of X i

F x

is in the domain of attraction of a Frechet type GEV iff

1 F x 1/
lim
=
x 1F x
for all

polynomial decay in the tail of

F x

GEV Types of Distribution

Conditions to the sample {X 1 , , X n } from which the maxima are drawn

Xi

must be independently identically distributed (i.i.d.)

Let

F x be the distribution of X i

F x

is in the domain of attraction of a Weibull type GEV iff

there exists

with

F F =1

and

1
1
1/
lim 1F F
1F F =
x
x
x

for all

F x

has a finite upper end point

GEV return level

Of interest often is return level

zm

value expected every m observation (block maxima)

1
Prob y z m =1G yz m =
m
calculated using invers distribution function (quantile function)

z m = G 1 = 1log 11/m

m
1

z m = log log 11/ m

can be estimated empirically as the

1
1
1

z m = G 1 = inf y F y
m
m

Block Maxima

Maximum daily precipitation for Nov-March (Winter) and May-Sept (Summer)

Summer
Winter

Dresden

Block Maxima

Block Maxima

Peak over Threshold

Peak over Threshold (POT)

{ X i uX i u}
very large threshold u
follow a Generalized Pareto Distribution (GPD)
Daily precipitation for Nov-March (green) and May-Sept (red)

Dresden

Peak over Threshold

Peak over Threshold (POT)
The distribution of

Y i := X i uX i u

exceedances over large threshold u

are asymptotically distributed following a
Generalized Pareto Distribution (GPD)
1/

y
H yX i u=11
u
two parameters

scale parameter
shape parameter

advantage: more efficient use of data

disadvantage: how to choose threshold not evident

POT Types of Distribution

GDP has same 3 types as GEV depending on shape parameter
Gumbel

=0

y
H y =1exp
u
exponential tail

Pareto (Frchet)

1H y~c y1/
polynomial tail behavior
Weibull

has upper end point

u
F =

Peak over Threshold

POT: threshold u = 15mm

Poisson Point GPD Process

Poisson point GPD process with intensity
1/

yu
A=t 2t 1 [1
]
u

on A=t 1, t 2 y ,

Estimation and Uncertainty

General concept of estimating parameters from a sample

{ y i } , i=1, , n
Maximum Likelihood (ML) Method
Assume

{ y i } are draw from a GEV (GPD,...) with unknown parameters

y i ~F y , ,

and PDF

f y , ,= F ' y , ,

Maximum Likelihood Method

The likelihood L of the sample is then
n

L , ,= i =1 f y i , ,
It is easier to minimize the negative logarithm of the likelihood
n

l , ,=i=1 log f y i , ,
in general there is no analytical solution for the minimum
with respect to the parameters
Minimize using numerical algorithms.
The estimates

, ,

maximize the likelihood of the data.

Profile log Likelihood

To Take Home
There exists a well elaborated statistical theory for extreme
values.
It applies to (almost) all (univariate) extremal problems.
EVT: extremes from a very large domain of stochastic processes
follow one of the three types: Gumbel, Frechet/Pareto, or Weibull

Only those three types characterize the behavior of extremes!

Note: Data need to be in the asymptotic limit of a EVD!

References
Coles, S (2001): An Introduction to Statistical Modeling of Extreme
Values. Springer Series in Statistics. Springer Verlag London. 208p
Beirlant, J; Y. Goegebeur; J. Segers; J. Teugels (2005): Statistics of
Extremes. Theory and Applications. John Wiley & Sons Ltd. 490p
Embrechts, Kppelberg, Mikosch (1997): Modelling Extremal Events for
Insurance and Finance. Springer Verlag Heidelberg.648p
Gumbel, E.J. (1958): Statistics of Extremes. (Dover Publication, New York
2004)
R Development Core Team (2003): R: A language and environment for
statistical computing, available at http://www.R-project.org
The evd and ismev Packages by Alec Stephenson and Stuard Coles